Geometric Algebra Representation and Ensemble Action Classification Method for 3D Skeleton Orientation Data

In this paper, we propose a novel human body posture representation based on Geometric Algebra to extract the angles and orientations of the most informative body joints to describe human body postures. As a motion usually consists of a number of postures, which are different even in the same type of motion. We treat the postures of a motion independently. For each posture, a new Geometric Algebra based skeleton posture descriptor is used to construct the feature vectors as the input for the Support Vector Machine classifier to decide its motion type. To get the type of the whole motion, we choose the most frequent class from the sequence of predictions of the motion postures using a simple voting scheme. We have tested the method on a public benchmark SYSU-3D-HIO and an in-house dataset of human exercises. The results have demonstrated the effectiveness of our method.